Research On Small Object Detection Problems In Pictures

Object detection is the very important research area in the field of computer vision.It is widely used in medical image analysis,navigation,defense systems,video tracking,video surveillance,unmanned aerial vehicle,autonomous driving and so on.However,the research on small obj ects detection is very limited.Most of the research is aimed at the large objects in the PASCAL VOC library,and ignores the detection of small obj ects which are occupied a small proportion of the image so that it is mor difficult to detect and more easily generates aggregation.This paper focuses on the problem of low accuracy and slow recognition speed in the small objects detection.Based on the image processing algorithm of deep learning,this paper proposes a convolutional neural network model based on Faster R-CNN which integrates with the flexible fusion of context information method and the effective dimension reduction of feature map method.The algorithm is fully experimented and evaluated.Firstly,a more flexible context information fusion approach is introduced into the Faster R-CNN model,which solves the problem that the small objects features extracted by the convolutional neural network are weak.Meanwhile,our approach effectively enhances the characteristics of the small objects features expression without introducing classification regression error;Secondly,there are three approaches which are respectively adaptive learning of small objects features extracted by the convolutional neural backbone network,different weights for each feature channel,and effective reduction of channel dimensions.They not only improve the accuracy of small obj ects detection,but also improve the speed of small obj ects recognition;Finally,the trained model based on optimization techniques of deep learning,such as online hard example mining and data arguement technology on training data set,improves the effectiveness of model training so that small object detection accuracy and speed have both improved.Experiments show that the small objects detection model proposed in this paper improves the accuracy of small obj ects detection by nearly 8%,improves the detection speed of small objects by nearly 20fps,and significantly reduces the computational complexity of convolutional neural networks.This paper applies the improved model to the real-time scenes of vehicle and pedestrian detection.We tunes the accuracy and speed of small object detection model and realizes the real-time detection system of long-distance vehicles and pedestrians.